PL seminar - correction reminder

[Shai Rubin <shai@xxxxxxxxxxx>]

All,

Due to a conflict with another talk, this week PL-seminar is rescheduled
to 2:30p, room 2310, (Thu 12/7/00).

Shai


---------- Forwarded message ----------
Date: Wed, 29 Nov 2000 23:35:23 -0600 (CST)
From: Shai Rubin <shai@xxxxxxxxxxx>
To: "CS Dept. Talks" <colloq@xxxxxxxxxxx>
Cc: Denis Gopan <gopan@xxxxxxxxxxx>

-N PL Seminar
-S Shai Rubin
-F UW - Madison
-T A Profile-Feedback Framework for Compiler Memory Optimization
-D 12/7/00
-W Thursday
-M 4:00
-P 2310 CS
-A
.pp
As the gap between the CPU and memory speeds increases, we anticipate that
memory optimization will become a dominant compiler component.    
.pp
This talk will first review the difficulties encounter when trying to
optimize program memory performance (e.g. cache misses, page faults).
First, such an optimization is global in nature, it requires whole program
analysis. Second, it deals with the dynamic program behavior. These two
properties call for profile based optimization. Furthermore, such
optimization effectiveness is based on the accuracy of the profile. Hence,
various precision profiles might be needed before the desired improvement
is achieved. Third, unlike profile-driven inlining and other traditional
optimizations, profile alone cannot make good decisions. The optimization
must be tuned by the impractical iterative process of instrumenting,
editing, compiling and executing the program. 
.pp
Next, we present a framework that overcomes these difficulties. The
framework stores the address trace in a tightly compressed form on which
profile summaries and other performance-analysis metrics (such as cache
miss rates) can be computed very efficiently. The framework enables
efficient analysis of the whole program memory behavior, efficient
generation of multiple profiles, and efficient iterative process of
tuning and evaluating the optimization under consideration. Finally, we
present a case study to show how the framework was used to analyze and
improve virtual-memory performance. We used highly accurate profile
information to direct custom-memory allocator to perform memory aware
objecst placement.
.pp
This is a joint work with Dr. Trishul Chilimbi (MSR) and Prof. Ras Bodik
(UW).